[clang] [Clang] Fix handling of brace ellison when building deduction guides (PR #94889)

[Younan Zhang via cfe-commits]

================
@@ -1449,7 +1449,10 @@ void InitListChecker::CheckSubElementType(const InitializedEntity &Entity,
       //   dependent non-array type or an array type with a value-dependent
       //   bound
       assert(AggrDeductionCandidateParamTypes);
-      if (!isa_and_nonnull<ConstantArrayType>(
+      // Don't consider the brace elision if the initializer is a
+      // braced-init-list.
+      if (isa<InitListExpr, DesignatedInitExpr>(expr) ||
----------------
zyn0217 wrote:

Thanks for the writeup! This matches what I thought.
A few more comments:

> as the e1 has a dependent type we can't perform semantic checks to see whether the T t[2] can be initialized by x1, we always assume true here. 

Yeah. For the case above, and as per [P2081](https://www.open-std.org/jtc1/sc22/wg21/docs/papers/2020/p2082r1.pdf)'s purposes, brace elision should be considered on arrays (possibly type-dependent, but the bound should be non-dependent), and what we used to do here is to always fall through to the brace elision logic below regardless of the presence of braces within the corresponding initializer `e`. And this results in issues of not having proper deduction guides.

> This has an effect that we will generate the deduction guide even for invalid case e.g. x1 is {1, 2, 3} ...

Yep, we don't guard against invalid initializers here because the `InitListChecker` is merely employed to generate "candidate" parameters. (where we set `VerifyOnly` to true)

https://github.com/llvm/llvm-project/pull/94889